The Rental Era Is Ending. The Next AI Moat Is Owning Your Own Electrons.

A year ago the smart take was that compute had stopped being a moat. GPUs were getting easier to rent by the hour, and we argued that ourselves: the frontier was turning into a spot market where anyone with a credit card could buy training runs.

That call is aging in an interesting way. The biggest labs are now doing the opposite of renting.

This week OpenAI committed roughly $20 billion to build its first self-developed data center campus, a 3.2 gigawatt site in Effingham County, Georgia, tied to a long-term power and tax agreement with the local utility. Not a lease. Not a cloud reservation. Land, substations, and a power contract the company controls.

**Why the pendulum swung back**

The reason is that the bottleneck moved. When chips were scarce, whoever could source GPUs won. Now the scarce thing is power and the permission to use it. A gigawatt of steady electricity, a grid interconnect, and a county that will approve the substation are harder to get than a rack of accelerators.

When the scarce input shifts from a component you buy to infrastructure you assemble, vertical integration comes back. You cannot rent your way to a 3.2 gigawatt guarantee that holds for a decade. You have to own it.

This is not only an American move. In Korea, LG CNS and local investors just put up first-round financing for a smart data center in Namyangju, backed by equity and land acquisition rather than a hyperscaler lease. Different market, same logic: secure the physical base first.

**What this means for everyone who is not OpenAI**

Two things follow, and they point in opposite directions.

For the largest labs, owning power becomes a durable advantage that capital alone cannot copy quickly. A rival can raise $20 billion in a weekend. It cannot conjure a permitted 3.2 gigawatt site and a signed utility contract on the same timeline. The moat is measured in interconnection queues, not model weights.

For everyone else, the rental market that we described does not disappear, it becomes the layer you live in. Startups and even mid-size labs will keep buying compute by the hour, and that is fine, because their edge was never going to be owning power plants. Their job is to be the most efficient tenant, not the landlord.

The risk sits in the middle. A company big enough to feel it should own infrastructure, but not big enough to actually secure gigawatts, can burn enormous capital trying to straddle both. That is where this cycle will produce its most expensive mistakes.

The bigger picture is that AI capacity is turning into something closer to national infrastructure, which is why countries are starting to form alliances around it rather than leaving it to companies.

**What to do with this**

If you are placing bets in AI, stop treating "has raised a lot of money" as the signal for the infrastructure layer. Track two things instead. First, who has actually signed power and interconnection deals, not just announced spending, because the announcement is cheap and the megawatts are not. Second, watch whether a company is trying to own infrastructure it is too small to secure, that straddle is the tell for capital that will not come back.

And if you run a smaller AI company, resist the vertical-integration envy. The winning move for you is to be the sharpest tenant in a market that just got more, not less, competitive to rent from.